AI tools like ChatGPT are built on mass copyright infringement
theglobeandmail.com
theglobeandmail.com
Artists making copies of famous pictures is a standard part of artist training. Tribute bands are a thing. Elvis impersonators are a mini-industry. The lawsuits so far tend to involve "passing off" new work as someone else's work, as with that ML-generated popular song passed off as a new work by a major artist. That's not a copyright issue. That's ordinary fraud.
The bitching from writers and musicians reflects not that they're being ripped off, but that they're being out-produced. Authors didn't expect to be in the position of John Henry vs. the steam hammer. Now they are.
Most modern neural nets, and certainly LLMs, are overparameterized, which means they are completely capable of regurgitating exact training instances. They can and do repeat copyrighted training data verbatim. It's absolutely copyright infringement when that happens.
I anticipate AI researchers will soon come up with a way to ensure that outputs are a sufficient distance from all training inputs -- to essentially "block" direct regurgitation as valid output, with a tunable "max similarity" set to a safe distance. At that point we can start being more confident that our AI output is genuinely transformative and safe from (valid!) lawsuit.
> The bitching from writers and musicians reflects not that they're being ripped off, but that they're being out-produced.
This comes across as quite heartless. It will be interesting to see whether your view on this changes when AI starts making obsolete whatever you have spent your life specializing in.
EDIT: We should welcome automation and destruction of work where a machine can do it better. We fear it because of a broken socioeconomic system. Agree we should probably put breaks on anything so disruptive unless people impacted have some safeguards, because if you don't, what good is progress for progress' sake? It then becomes no different than bullshit jobs.
(e: and his edit says it as eloquently as I can!)
The problem is not automation... Automation is great. It frees up people's time to do more activities. The problem is how we distribute the spoils of that freed time.
Telephone operators. Typing pools. Grooms. TV repair technicians. Weavers
To this list we may be adding law clerks. Drafters. Copy writers.
This is an ancient process. Some societies and cultures cope and maintain a civil society, some don't
So do not play "whack'a'mole", campaign for social change
While you're at it, take computers always from people. They automate way too much stuff.
Btw, I am in one of those careers that people think is at the risk of being automated away. I don't really care about it. If I become obsolete, maybe my skills were not that useful. I'll try working in something else.
And if all jobs are automated away, I'll just go starve with the rest of the peasantry. The last to starve will be the first to suffocate.
Things will look very similar because there are only so many things that sound good to the mass appeal.
> I anticipate AI researchers will soon come up with a way to ensure that outputs are a sufficient distance from all training inputs
This is already a thing.
A better example is that Vanilla Ice had to pay royalties to Queen and David Bowie for using the bass hook from Under Pressure in Ice Ice Baby.
They're really not that similar. The Marvin Gaye family lawsuits are absolutely frivolous and a stain on his legacy. Suing over the same chord progression (ignoring tons of small variations that actually matter) would be like suing an author because they also had a character who was an angsty teenager with overbearing parents. Suing over a mostly-kinda-similar riff is like suing an author because they also used the sentence "Terrified, he ran." The songs only sound similar when you ignore all elements of the whole song and focus exactly on a few ways in which they overlap out of context.
They're nothing like GitHub CoPilot literally outputting comments that include copyright notifications from their original author. It's much more like suing someone who directly samples your song as a core element of their own (the Vanilla Ice example in the other comment).
Alright, but how do you define this when talking about most of modern "pop" music (using a broad definition that also includes things like techno or rock sub-genres)?
Maybe even only consider what's commercially relevant.
Copyright is famously about authorship as an ulterior value, governing and including other values such as monetization.
What about this aspect?
Creativity is also often about subtlety, e.g. polished songwriting and sound design.
Already without AI, rent-seeking is prevalent in popular music, maybe it's even an integral part?
For a stupid example of reusing themes in pop music, take this especially spicy example: https://en.wikipedia.org/wiki/Kernkraft_400
Or listen to the radio maybe. As a 90s and 2000s dance music lover, there's plenty of rehashing old content, sometimes licensed, other times it's just what music is.
Songs are memes after all
My brain can do this too with music. In my 20's I used to be able to hear songs crystal clear in my head as if they were coming from the stereo itself.
Did my brain copy the song (yes!). Did it receive copyright approval from the artist (no!).
No they're not.
GPT-3: Model size: 175 GB, training set size 300 billion tokens.
GPT-4: Not disclosed.
LLaMA: Model size: 65GB, training set size, 1 trillion tokens.
These things condense the data. They don't just index it, like a search engine.
If it was just some code, included into a compiled EXE, without permission - it would be a major violation. But including books inside of an AI model - it is different (not really).
https://www.copyright.gov/title17/92chap1.html#106
106. Exclusive rights in copyrighted works
Subject to sections 107 through 122, the owner of copyright under this title has the exclusive rights to do and to authorize any of the following:
(2) to prepare derivative works based upon the copyrighted work;
https://www.copyright.gov/title17/92chap1.html#107
107. Limitations on exclusive rights: Fair use41
Notwithstanding the provisions of sections 106 and 106A, the fair use of a copyrighted work, including such use by reproduction in copies or phonorecords or by any other means specified by that section, for purposes such as criticism, comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research, is not an infringement of copyright. In determining whether the use made of a work in any particular case is a fair use the factors to be considered shall include—
(1) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;
(2) the nature of the copyrighted work;
(3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
(4) the effect of the use upon the potential market for or value of the copyrighted work.
The fact that a work is unpublished shall not itself bar a finding of fair use if such finding is made upon consideration of all the above factors.
===
What will the courts decide? Who knows. It's probably going to be a fairly massive case regardless.
Copyright violation is generally on the output side of things, not the input. It's not a copyright violation to watch a movie.
I also would nit pick that a model is not BASED UPON a copywritten work, just because consuming the copywritten work changed the models parameters slightly. The model itself is not a derivative of any specific work. If one work represented over 50% of a model, it might be a different story. People are getting a little over focused on the word derivative and ignoring the rest of the sentence.
Copyright infringement is doing anything that violates a exclusive right of a copyright holder[1]. There are a bunch of exclusive rights one of which involves the output side. The Southern District of New York Federal court just sided with publishers against the Internet Archive specifically in deciding that scanning books was a copyright violation. [2]. Particularly of note is page 14 of the decision [3]
---
> The Publishers have established a prima facie case of copyright infringement. First, the Publishers hold exclusive publishing rights in the Works in Suit pursuant to 17 U.S.C. § 106, and the Works were timely registered with the Copyright Office. Pls.’ 56.1 ¶¶ 210-214. Second, IA copied the entire Works in Suit without the Publishers’ permission. See id. ¶¶ 242-245. Specifically, IA does not dispute that it violated the Publishers’ reproduction rights, by creating copies of the Works in Suit, 17 U.S.C. § 106(1); the Publishers’ rights to prepare derivative works, by “recasting” the Publishers’ print books into ebooks, id. § 106(2); the Publishers’ distribution rights, by distributing ebook copies of the Works in Suit to IA’s users, id. § 106(3); the Publishers’ public performance rights, through the “read aloud” function on IA’s Website, id. § 106(4); and the Publishers’ display rights, by showing the Works in Suit to users through IA’s in-browser viewer, id. § 106(5).5
---
The generation of the model literally requires a step which converts the source material into tokens. That step is infringing. It's no different to taking photocopies of a photo, or whipping my phone out during a movie at the theater to record. Whether it's infringing is not the issue. It is. It's whether the act is fair use that is the problem.
I guess what you're trying to argue here is that using the resulting model is non-infringing. That's a whole 'nother thing I guess.
---
[1]: https://www.copyright.gov/title17/92chap5.html
[2]: https://en.wikipedia.org/wiki/Hachette_v._Internet_Archive
Had they archived the books and not lent out unlimited copies, nothing would have come of the case.
Obviously LLM outputs are transformative, so this argument falls completely flat. As the writer is a copyright lawyer, it's hard to conclude anything other than they are knowingly lying, or at minimum wishcasting what they want the law to say instead of what it does say.
I think the misconception stems from the laymen understanding of copyright clipping off the last part of that sentence so it's just "Copyright infringement is when you take copyrighted work".
Proof of the success of industry campaigns to vilify things like taping broadcast television.
What if Greg isn't an llm and he's your browser cache? Are you still infringing copyright?
Kim Peek is Greg and the difference between him and and AI is the text above.
> Should you have to pay royalties to everyone whose content you ask Greg to look at?
If Greg talks so fast that he's distributing millions of these copies around the world, for money, then yes, of course he's infringing.
> What if Greg isn't an llm and he's your browser cache?
My browser cache is not a distribution mechanism. It's for my personal use. I'm not infringing on copyright if I keep books in my personal library. I am if I'm copying them millions of times and giving others access to that library for money. If I downloaded a bunch of paywalled content and then uploaded my browser cache to SomePirateSite.com, for money, then yes, I'm infringing.
Why do you think these are "gotcha" questions? This is pretty straightforward stuff, and nowhere does it prove that LLMs are not infringing.
So you mean Google, right?
Thats an interesting notion but isnt this just as true of any generative logic, provided the size of the set of results is larger than the code size? Like a random number generator. Can you compress infinity?
Distribution isn't necessarily a required element of infringement. Just the creation of a derivative work can be enough.
“Transformative” is one consideration in one of four factors for fair use, so that's not right on the exception, and infringement happens when any exclusive right (literal copying, creating a derivative work, public performance, etc) is done with neither permission nor an exception such as Fair Use, not just at distribution.
It goes much further. Loading copyrighted work into a computer's memory, for example, is copying and can infringe. There are 30 years of precedent here. See https://en.wikipedia.org/wiki/MAI_Systems_Corp._v._Peak_Comp....
Pointing out that a major advancement would be foreclosed by a proposed interpretation of the scope of copyright law is an argument that that interpretation is either an incorrect interpretation of the statute or an area where the statute conflicts with its Constitutional purpose.
Of course, the Globe and Mail is Canadian, but, is Canadian law the applicable governing law here?
1. The benefactors of these models can offer to pay a small fee to the copyright holders for inclusion of their work in the training set. Given that these models have the potential to be literally the most monetarily valuable products ever created in the history of mankind, paying the people who allowed them to exist doesn't seem like an unreasonable ask to me. Obviously people who refused would not have their works included.
2. The model can be trained on public works. Sometimes this is plenty, sometimes not.
3. Hordes of human beings can be paid to contribute to the training data set, Mechanical-Turk-style or just via direct employment. This is already necessary to curate and polish the input data.
None of these seem like outlandish requests for something that will eventually be worth trillions of dollars per year.
Consider the opposite: we give an exception to giant corporations and say "yeah feel free to just take whatever copyrighted work and use it however with no compensation." Artists and authors are being replaced by these outputs. What motivation do they have to create, knowing that their work will be immediately stolen by giant corporations for use in building the tool that makes them obsolete? How is that promoting “the progress of science and useful arts”?
I would like to see more high-speed railway exist. By your argument, land ownership is problematic because it prevents cool things like high-speed rail, and therefore we should just plow on through everyone's land for the greater good without compensation. I say: Eminent Domain is a useful last resort to make cool things happen, but at least compensate the people you're trampling on.
I do not understand. How can you justify that for publicly available text (like this?)
Why is a LLM learning from text a misuse yet you learning from text not? (See my assumption?)
There is no evidence (?) that these LLMs are trained on anything other than public Internet.
The "land ownership " analogy is unhelpful. Physical objects are entirely different from intellectual objects
FWIW, the author appears to be the founder or former founder of Licensy - you can guess what it's about, so I wouldn't take the article as a neutral legal opinion anyway.
Copyright's current state might be excessive, but doing away with it entirely is almost surely not the answer.
a) a copyright holder’s ability to monetize their work would not be impacted, or
b) that all the people that currently make their living creating copyrighted works, would continue to invest their time and money to produce their works regardless of their ability to be compensated
I personally find both premises to be unsupportable.
But you also believe that the richness of culture would not diminish in any way.
I wish you'd explain what you think would happen.
(I can say that if copyright stopped existing my business would also immediately stop existing and I'd have to find a new career or a new country).
Who are you talking about? What is their living like? From what I can see writer's and musicians mostly live in poverty unless they are hobbyists. Like me.
I produce audio art (eg https://m.youtube.com/watch?v=jrAAPbg81ks). No thought, ever, to copyright.
And yes plenty of people try unsuccessfully to make a career from their art or just do it for fun as a hobby.
Agreed that 'current state' is excessive. It's a shame that long out-of-print works (which are direct evidence of little or no commercial value to authors/publishers) have not become freely available since digitization came along. (Paying $35 for a digital copy of an ancient, but important science paper, for example.) OOP-status is pretty damned easy to verify today.
Zero
What a good idea!
Some more space for new art, rather than the warmed up recycled shirt we get now
The problem with stuff that isn’t available online is the cost of getting it into a form that can be ingested.
This may be fair in the short term, but on a longer timescale, it leaves us worse off unless LLMs master reasoning about and being creative with things not in their training set. Possible, but a crapshoot right now.
Copyright law cannot be allowed to stand in the way of this. If it does, break it. It's that important.
UBI is always possible. The achievable minimum support level might be higher with this, but also, with this occurring without UBI existing first, so will the concentration of economic and political power in the hands of entrenched interests who would want to prevent UBI (since it would be redistributing resources outward from them.)
Couple "the cruelty is the point" with figuring out exactly who will be giving up resources for UBI, and you have a very long road to hoe to make that a reality.
UBI is already technologically possible. Yes, as automation becomes even more powerful and abundant, the feasible payouts from UBI will increase. That's never been the problem. The problem is that hundreds of millions of Americans (and a significant portion of everyone else in the world) would rather live in poverty as indentured servants to their rich masters than ever see those people benefit even a tiny bit from any government policy. Cruelty and suffering is literally what they want in the world; they fight tooth and nail to prevent anything that might uplift all of humanity (themselves included). They are deeply disturbed, horrible, irredeemable pieces of shit, and they're the majority.
And the very tiny minority of people who could ever do anything about this are exactly the ones who are poised to take all the benefits from automation and AI. I'm not scared of AI. I'm scared of humans. And very soon, the humans who already own everything are going to have no need for the rest of us.
After that we could explore other economic models that make sense in a world of internet and AI, instead of blindly sticking with the system we designed for the world of the printing press.
1. Create enough art in the same restricted domain, and some works will happen to look similar to others.
2. Use aspects of a work in the process of creating another. Without enough process, the result will resemble the original.
LLMs pretend to do #1: They first model a collection of works - called the training corpus - to implicitly generate a domain. Everything that is "hallucinated" by that model is restricted to that domain.
The entire goal of an LLM is to resemble the training corpus. The only interesting part is "in what way" that resemblance appears.
For example, an infamous paper illustrated that after being trained against valid Othello games (written as sequential piece positions), GPT can hallucinate a valid Othello game move. Does this mean that the game moves are novel art? Of course not! They are the result of the training corpus being modeled through a transformer. This is method #2 masquerading as method #1.
I'll entertain arguments from either side of the infringement debate, but this opinion piece reads more like a reactionary defense of copyright rather than a measured response. These damages are... not that significant. It feels like Big Tech swooping in to defend Open Source not because they care about freedom, but because without it they would have no software. But instead of copyleft, it's copyright. Hmm.
> the net cost of copying data is effectively nothing.
Since the cost of including the author and the link to original work is so cheap I dont see why not have the courtesy of adding that info. They're are already doing the work for free, recognition is the least you could pay in respect for it.
Granted, sometimes things are a complex tree of authorship transformations and it's an interesting question of how far back you d go for that but sometimes its pretty cut and dry too.
Look how justice responded when MZ was sued: first, by his university, then by the justice system. None did anything.
I'll let you just imagine what would have happened if MZ was of another ethnicity. Well, not only ethnicity but the other factor is a taboo even here.
If you don't have it, your competitors will, and your people will use the competitors software anyway.
What about musician training? Movie directors?
(a) is it so similar to the original as to be considered derivative?
(b) did the author have any exposure at all to the original?
Movies, music, books: it doesn't matter. Lots of people are focusing on lots of red herrings in this thread. If the AI outputs its own copyrighted training data verbatim, it's copyright infringement, full stop.
Or at least the last time I checked nobody had forked my brain on GitHub or uploaded me to Sourceforge.
Probably the same goes for open content licenses.
On one hand, that makes her an expert in the legal aspects of this case. On the other hand, generated content presents a substantial threat to her current income stream.
Seems like the author was pretty clear to me about their history.
The problem is that under current copyright law, having to purchase a distribution license for everything an AI learns in its training data would be cost prohibitive, to say the least, which also stagnates art and innovation.
I imagine a future where much of human effort is ultimately directed towards creating ground-truth data for a generative AI, whether that be text, pictures, art, or other media. I have a writer friend for whom this is happening already. We need to somehow incentivize ground-truth collection, and we don't have a way to do that right now apart from wages, which IMO is not sufficient.
To me, the question is not whether or not AI violates copyright, because we're clearly in new territory here. The question is, how do we properly compensate folks who are creating that ground-truth data? Do we need to restructure our economy from the ground up, or is there a path from the current capitalist economic model to doing this?